Successful Companies vs Closed and Acquired: Data Analysis and Logistic regression model

Preamble:

This report presents an in-depth analysis of venture capital investments in Latin American startups, with the goal of identifying key characteristics that distinguish successful companies from those that fail.

The analysis is carried out in two stages: the first stage involves an exploratory data study of various Latin American startups across numerous metrics, while the second stage focuses on the development of a Logistic Regression model designed to predict the probability of a company’s success or failure.

For the purpose of this analysis, we define success in the following way:

Successful companies are those that have secured at least one funding round exceeding 20 million dollars.

In contrast, unsuccessful companies are identified as those that have either shut down or have been acquired as a means of survival.

# Libraries to use. We load them quietly
packages <- c("tidyverse", "googlesheets4", "glmnet", "caret", "ggthemes", "ggpubr", "stargazer",
              "psych","summarytools","kableExtra","scales","reshape2","car", "forcats", "caTools",
              "tidymodels","pROC", "ModelMetrics","oddsratio","epiDisplay","gridExtra")

invisible(lapply(packages, function(x) {
  suppressPackageStartupMessages(library(x, character.only = TRUE, quietly = TRUE))}))

The data for this analysis is sourced from the Pitchbook platform. Now, let’s begin by processing the raw dataset.

raw_data_vc  <- sheet_data_invest %>%
  #Rename Variables
  rename(company= "Company",
         vertical= "Vertical General",
         status = "Status",
         hq = "HQ Company",
         deal_type = "Deal Type",
         team_composition = "Composición del equipo",
         founding_year ="Founding year (company)",
         funding_year = "Funding Year",
         funding_usd = "USD",
         employees = "Employees",
         total_raised= "Total Raised",
         funding_date = "Funding Date",
         n_investors = "Number of investors",
         clasf_unicorn= `New clasf unicorn`,
         exit_date = `Exit year`,
         team = `Composición del equipo`,
         hq_inv = `Headquarters of Lead Investor`) %>% 
  # Group by company name
  group_by(company) %>%
  # Mutate different variables. Here I'm correcting some data capture errors.
  mutate(funding_date = as.Date(funding_date, format = "%Y-%m-%d"),
         founding_year =as.Date(paste(founding_year, "-01-01", sep = ""), format = "%Y-%m-%d"),
         exit_date =as.Date(paste(exit_date, "-01-01", sep = ""), format = "%Y-%m-%d"),
         end_date = as.Date(ifelse(status == "Closed", as.character(max(funding_date)),
                                   ifelse(status == "Acquired", as.character(exit_date), NA)), format = "%Y-%m-%d"),
         vertical = ifelse(vertical %in% c("Insurtech", "InsurTech"), "Insurtech", vertical),
         deal_type = ifelse(deal_type %in% c("Line of Credit", "Line of credit"), "Line of credit", deal_type),
         team = factor(ifelse(team %in% c("Equipo femenino", "Equipo Femenino"), "Equipo femenino", 
                              ifelse(team %in% c("Equipo Mixto","Equipo mixto"),"Equipo mixto",team)),
                       levels = c("Equipo femenino","Equipo masculino","Equipo mixto",
                                  "Un solo hombre","Una sola mujer")),
         status = factor(ifelse(status %in% c("acquired", "Acquired"), "Acquired", status),
                         levels = c("Acquired","Closed","Operating")),
         clasf_unicorn = factor(ifelse(clasf_unicorn %in% c("Was acquired","Was Acquired"), "Was Acquired", clasf_unicorn),
                                levels = c("IPO","Rest of Companies","Soonicorn","Unicorn","Was Acquired")),
         # Calculate new variables based on existing ones. For example, get median funding by company
         median_funding = median(funding_usd, na.rm= T),
         total_investors = sum(n_investors),
         funding_rounds = n(),
         foreign_funding_all = if_else(hq_inv== hq,"No","Yes"),
         foreign_funding_all = factor(foreign_funding_all,levels = c("No","Yes")),
         foreign_funding = ifelse(
           hq != hq_inv &
             !(hq_inv %in% c("Argentina", "Brazil", "Chile", 
                             "Colombia", "Ecuador", "Mexico", 
                             "Peru", "Venezuela")),
           "Yes",
           "No"),
         same_country_funding = ifelse(hq==hq_inv,"Yes","No"),
         regional_funding = ifelse(hq %in% c("Argentina", "Brazil", "Chile", 
                                             "Colombia", "Ecuador", "Mexico", 
                                             "Peru", "Venezuela") &
                                     hq_inv %in% c("Argentina", "Brazil", 
                                                   "Chile", "Colombia", "Ecuador", 
                                                   "Mexico", "Peru", "Venezuela"), "Yes", "No"),
         # Here I'm manually correcting some data capture errors
         first_funding_round = case_when(
           company == "Brex" ~ as.Date("2019-04-16", format = "%Y-%m-%d"),
           company == "Hackmetrix" ~ as.Date("2020-07-01", format = "%Y-%m-%d"),
           company == "Hash (Financial Software)" ~ as.Date("2017-01-05", format = "%Y-%m-%d"),
           company == "Inventa (B2B Marketplace)" ~ as.Date("2021-08-17", format = "%Y-%m-%d"),
           company == "Kavak" ~ as.Date("2016-08-01", format = "%Y-%m-%d"),
           company == "Nexu" ~ as.Date("2014-01-01", format = "%Y-%m-%d"),
           company == "Solfacil" ~ as.Date("2018-01-01", format = "%Y-%m-%d"),
           TRUE ~ min(funding_date)),
         last_funding_round = case_when(
           company == "Brex" ~ as.Date("2019-12-11", format = "%Y-%m-%d"),
           company == "Hackmetrix" ~ as.Date("2021-11-28", format = "%Y-%m-%d"),
           company == "Hash (Financial Software)" ~ as.Date("2021-10-20", format = "%Y-%m-%d"),
           company == "Inventa (B2B Marketplace)" ~ as.Date("2022-01-23", format = "%Y-%m-%d"),
           company == "Kavak" ~ as.Date("2022-09-20", format = "%Y-%m-%d"),
           company == "Nexu" ~ as.Date("2022-01-26", format = "%Y-%m-%d"),
           company == "Solfacil" ~ as.Date("2022-09-22", format = "%Y-%m-%d"),
           TRUE ~ max(funding_date)),
         months_to_first_round = as.numeric(round(difftime(first_funding_round,founding_year, units = "days"))/30, digits =1),
         months_between_first_and_last_round = as.numeric(round(difftime(last_funding_round,
                                                                         first_funding_round, units = "days"))/30, digits=1),
         months_between_rounds = c(NA, diff(as.numeric(difftime(funding_date, lag(funding_date), units = "days"))/30)),
         team = case_when(
           team == "Equipo femenino" ~ "Co-Founder Team",
           team == "Equipo masculino" ~ "Co-Founder Team",
           team == "Equipo mixto" ~ "Co-Founder Team",
           team == "Un solo hombre" ~ "Solo Founder",
           team == "Una sola mujer" ~ "Solo Founder",
           TRUE  ~ team),
         team = factor(team,levels = c("Co-Founder Team","Solo Founder")),
         avg_months_between_rounds = mean(months_between_rounds, na.rm = TRUE)) %>% 
  # Filter out NAs and absurd variables, such as a company having 0 investors or negative total raised
  filter(complete.cases(funding_usd,total_raised, n_investors, founding_year, employees, funding_date),
         !grepl("Undisclosed - VC", deal_type),
         n_investors >0,
         total_raised >0,
         as.numeric(format(founding_year, "%Y")) >= 2013) %>% 
  # Company classification starts from this point.
  # Companies classified as Unicorns or Soonicorns are considered to be those who have gotten a +20M USD Funding round
  # I'm only considering companies that closed or were acquired from 2018 to beyond.
  filter(clasf_unicorn%in% c("Unicorn","Soonicorn")|status%in% c("Closed","Acquired") 
         &  as.numeric(format(end_date, "%Y")) >= 2018) %>% 
  # Successful companies will be those with a Unicorn or Soonicorn status
  # Not Successful companies will be those that were closed or acquired
  mutate(success = ifelse(clasf_unicorn %in% c("Unicorn", "Soonicorn"), "Successful",
                          ifelse(status%in% c("Closed","Acquired"),"Not successful",NA)),
         success = ifelse(company %in% c("Auth0", "Checkars", "Cornershop","Daki"), "Successful", success),
         success = factor(success,levels = c("Successful","Not successful"))) %>% 
  dplyr::select(company,vertical, hq, hq_inv,success, founding_year,end_date,foreign_funding_all, same_country_funding,
         median_funding, funding_usd,
         regional_funding, employees, team, months_to_first_round,
         months_between_first_and_last_round, avg_months_between_rounds, 
         total_raised, funding_rounds, n_investors) %>% 
  distinct(company, .keep_all = T)

This data was obtained via a LinkedIn Scrapper, let’s clean and join both the investments data with the founder data

raw_data_scraper <- sheet_data_scraper %>% 
  # Rename Variables
  rename(company = `Nombre de la compañía del mapeo`,
         exp_years = "Años de Experiencia",
         entrep_exp_years = "Años de experiencia como emprendedor",
         age = "Edad Actual",
         entrep_age= "Edad de emprendimiento",
         n_entrep_founded = "Numero de emprendimientos Fundados",
         hundred_employees = "Worked on a +100 Company",
         status = "Status Unicorn (1=unicorn, 3=soonicorn, 0=other companies)",
         foreign_exp = "Foreign exp",
         cofounders = "No. de Cofounders",
         founders = "# founders",
         employments_before_entrep= "Empleos previos a emprender",
         business = Business,
         entrep = Entrepreneurship,
         finance = Financial,
         marketing = Marketing,
         maths_eng = `Maths & Engineering`,
         pmanag= `Project Management`,
         research = Research,
         sales = Sales,
         soft = `Soft Skills`,
         consulting_exp = Consulting_exp,
         founder_exp = Founder_exp,
         clevel_exp= cLevel_exp,
         field_of_study = `Undergraduate field of study`,
         exp_years_categ = `Categorización Años de experiencia`,
         foreign_exp = "Foreign exp") %>% 
  group_by(company) %>% 
  dplyr::select(company,exp_years, exp_years_categ,cofounders,
         entrep_exp_years,employments_before_entrep,
         age,entrep_age,n_entrep_founded,hundred_employees,
         business,entrep,finance,marketing,maths_eng,pmanag,
         research,sales,soft,manager_exp,clevel_exp,engineering_exp,
         sales_exp,finance_exp,human_capital_exp,consulting_exp,founder_exp, field_of_study, foreign_exp) %>%
  # Mutate and create new variables by company
  mutate(cofounders = dplyr::recode(cofounders,
                             "2 Cofounders" = 2,
                             "3 Cofounders" = 3,
                             "4 Cofounders" = 4,
                             "5 o más Cofounders" = 5,
                             "Single founder" = 1),
         exp_years = sum(exp_years),
         entrep_exp_years=sum(entrep_exp_years),
         employments_before_entrep = sum(employments_before_entrep),
         age = median(age),
         entrep_age = median(entrep_age),
         across(n_entrep_founded:founder_exp,sum)) %>% 
  filter(exp_years >= 0) %>% 
  # Keep a unique observation for each company
  distinct(company, .keep_all = T)

# Join founder data with the investments their company received.
joined_data_scraper <- left_join(raw_data_vc,raw_data_scraper, by = "company") %>% 
  mutate(cofounders = as.character(cofounders),
         cofounders = if_else(cofounders %in% c("4", "5"), "4+", cofounders)) %>% 
  filter(!is.na(cofounders)) %>% 
  group_by(cofounders) %>% 
  mutate(team = case_when(
    cofounders == 1 ~ "Solo Founder",
    TRUE ~ "Co-Founder Team"))

For confidentiality reasons, I can’t provide the complete data frame. However, I’m able to share a more condensed, high-level view of the data:

Figure 1.
Glimpse on Companies Data

company	hq	total_raised	months_between_first_and_last_round	team
NetLex	Brazil	53230000	32.06667	Solo Founder
Sofía	Mexico	31600000	10.43333	Co-Founder Team
La Plataforma	Chile	80000	19.50000	Co-Founder Team
DogHero	Brazil	12500000	31.10000	Co-Founder Team
byprice.com	Mexico	51760000	41.43333	Co-Founder Team

Part 1: Understanding Key Differences

First, let’s take a look at the median values. Given the non-normal distribution of the data, I’m focusing on examining the median values for each pertinent variable in the processed dataset.

Figure 2.
Summary Table: Median Values

Variable	Successful	Not successful
Months to first round	17	19
Months between first and last round	29	14
Funding rounds	4	2
Employees	174	11
Investors	3	1
Founding Year	2017	2016
End Year	NA	2020

The table reveals some intriguing findings:

- Successful companies typically secure their initial funding round two months ahead of unsuccessful ones.
- Successful companies usually have two additional funding rounds compared to unsuccessful ones.
- Unsuccessful companies often struggle to expand their workforce beyond 50 employees.
- The highest number of closures among unsuccessful companies occurred in 2020, likely influenced by the pandemic and issues accessing funding.

Now let’s plot some key differences

Figure 3. Team Composition Distribution

- While we shouldn’t equate correlation with causation, the graph does suggest an interesting pattern. Despite the fact that having co-founders doesn’t assure success, among co-founder companies, only 17% failed. While within solo founders, this percentage increases to 30% failed companies

Figure 4: Top Performing Industries

- E-commerce stands out as the most successful industry, with 84% of companies in this sector achieving success.
- Conversely, the AdTech/Media industry exhibits the highest failure rate, as 29% of companies within this sector have either shut down or been acquired.

Figure 5. Funding Locations

Another noteworthy observation pertains to the geographical origin of funding that entrepreneurs secure.

- Successful companies have 72% rate of getting a foreign funding round.
- On the other hand, unsuccessful companies get much more funding from local funds, with 52% receiving funds within their same country versus 28% for successful companies

Figure 6. Distribution of first entrepreneurship

- The graph displays the age distribution of their first venture for both teams.
- Solo founders tend to launch their first venture two years earlier than co-founder teams.
- The bubble size represents the number of employees for each age.
-It’s observed that for both teams, the longer their venture has been operating, the more employees they manage.

Figure 7. Correlation graph

- Both groups display a positive correlation: the more experience they have, the more investment rounds they secure.
- The correlation value for the team of solo founders is higher (0.33). This could suggest that solo founders with more experience may secure more investment rounds than co-founder teams with similar years of experience.
-The p-value is less than 0.05, indicating a significant correlation, unlikely to be caused by chance.

Figure 8. Skills Comparison

- There’s a 31% increase in individuals with product management skills in co-founder teams versus solo founders.
- Financial skills come second, with a difference of 23% in co-founder teams versus solo founders.
- Both groups excel in business operations, emphasizing the importance of managing ventures effectively, regardless of the number of founders.

Figure 9. Time it takes companies to move from one funding round to another

The timeframe a company takes to move forward to the next investment stages is another key element that investors consider. The graph shows the median duration a company typically takes to progress from an early to a later investment stage. A few interesting observations are:

- Successful companies usually secure a Series A investment within a year. In contrast, unsuccessful companies can take up to a year and a half.
- The shift to a Series B is challenging for all companies, but successful ones manage it in 13 months, while unsuccessful ones can take as long as 30 months — over twice as long.
- This trend continues into the transition to Series C, where unsuccessful companies require 27 months.
- Unsuccessful companies generally fail to reach a Series D or further rounds.
- Successful companies maintain a steady influx of investments, with an average interval of a year between rounds.

Part 2: Fitting a logistic regression model

Now, based on the previous analysis. I’m developing a logistic regression model.

set.seed(123)
# Select all numeric variables 
numeric_companies <- raw_data_vc %>% 
  # Mutate success as a number, where 1 = Success and 0 = Not Successful
  # Mutate team as a number, where 1 = Co-Founder Team and 0 = Solo Founder
  # Mutate foreign funding as a number, where 1 = Has received foreign funding and 0 = Has not received foreign funding
  mutate(success= as.numeric(if_else(success == "Successful",1,0)),
         team = as.numeric(if_else(team == "Co-Founder Team",1,0)),
         foreign_funding_all = as.numeric(if_else(foreign_funding_all== "Yes",1,0))) %>% 
  ungroup() %>% 
  select_if(is.numeric)

# Mutate variables as factors so the model captures the levels correctly
numeric_companies$success <- as.factor(numeric_companies$success)
numeric_companies$team <- as.factor(numeric_companies$team)
numeric_companies$foreign_funding_all <- as.factor(numeric_companies$foreign_funding_all)

# Split training and testing datasets respectively 
datos_split = initial_split(numeric_companies, prop = 0.7,strata = success)
datos_train = training(datos_split)
datos_test  = testing(datos_split)

# Fit the logistic regression model where the variable to predict is success of a company
logit <- glm(success ~ ., data = datos_train %>% dplyr::select(c(-avg_months_between_rounds,
                                                          -funding_usd,
                                                          -median_funding)), family = "binomial")

• Variable to predict will be Success of a company.

• 70% of the data will be assigned into training the model

• The results highlight statistical significance across three variables at a 95% confidence level.
  

• The model’s AIC Criteria is low, indicating its effectiveness.
  

• While it would be ideal to observe statistical significance across all regressors, the focus of this analysis is more on the model’s predictive capabilities and accuracy. Therefore, this aspect doesn’t undermine the analysis.
  

Logistic Regression Summary

	Dependent variable:
	company_success
Received Foreign Funding (1=Yes)1	-0.15
	(0.61)
Total Employees	0.01**
	(0.005)
Team Composition (1= Co-Founder Team)1	1.59**
	(0.78)
Months to achieve first round	-0.01
	(0.01)
Months between first and last round	-0.01
	(0.02)
Total Raised	0.0000***
	(0.00)
Total funding rounds	0.44
	(0.27)
Total Investors	0.29**
	(0.15)
Constant	-3.55***
	(1.02)
Observations	165
Log Likelihood	-38.57
Akaike Inf. Crit.	95.14
Note:	p<0.1; p<0.05; p<0.01

• The Variance Inflation Factor (VIF) for the regression is relatively low, particularly when we consider a benchmark of 5 as high.
  
• This suggests that there’s no substantial multicollinearity among the variables.
  

  VIF Values for each regressor

  	VIF Value
  Received Foreign Funding (1=Yes)	1.112138
  Total Employees	1.110112
  Team Composition (1= Co-Founder Team)	1.143243
  Months to achieve first round	1.170025
  Months between first and last round	2.437043
  Total Raised	1.106142
  Total funding rounds	2.558997
  Total Investors	1.042223

  
  
  

Confusion Matrices

• Comparing the model’s predictions on the training and testing datasets, there are minimal false positives and false negatives.
  
• This indicates that the model is performing effectively on both datasets.
  

Confusion Matrix for Testing Dataset

	Actual Failure	Actual Success
Predicted Failure	10	2
Predicted Success	4	57

Confusion Matrix for Training Dataset

	Actual Failure	Actual Success
Predicted Failure	27	11
Predicted Success	5	122

Model Performance Metrics

• The accuracy level is notably high for both datasets, exceeding 90%, with a variance of no more than 1% between them.
  
• The model provides accurate predictions across both these datasets which should give a very high confidence on predictions made.
  

Model performance metrics for testing dataset

Accuracy	Precision	Recall	F1_Score
0.9178082	0.9661017	0.9344262	0.95

Model performance metrics for training dataset

Accuracy	Precision	Recall	F1_Score
0.9030303	0.9172932	0.9606299	0.9384615

Area under the curve.

• To maximize the precision of predictions, I assessed the Area Under the Curve (AUC) for the model, in addition to conventional accuracy and F1 Score.
  
• The AUC values are impressively high, nearing 100%, and show negligible variation between the testing and training datasets.
  
• These results underscore the effectiveness of the logistic regression model, emphasizing its enhanced discrimination capability and predictive accuracy.
  
  

Predictions on real world data.

The following data frames contain real data from two companies.
Due to confidentiality obligations, I won’t disclose the names of these companies. However, this data is crucial for making predictions about their success.

Company A Data

Total Employees	Received Foreign Funding (1=Yes)	Team Composition (1= Co-Founder Team)	Months to achieve first round	Months between first and last round	Total Raised	Total funding rounds	Total Investors
109	1	1	12	16	17310000	3	4

Company B Data

Total Employees	Received Foreign Funding (0= No)	Team Composition (0= Solo Founder)	Months to achieve first round	Months between first and last round	Total Raised	Total funding rounds	Total Investors
26	1	0	24	36	10000000	2	2

Chance of being successful.

As seen in the provided data, there are significant differences between the two companies:

1. Company A has successfully expanded to over 50 employees, indicating consistent growth, while Company B hasn’t yet reached this milestone.

2. Both companies have attracted foreign investment.

3. Company A was launched by a team of co-founders, whereas Company B was started by a solo entrepreneur. As previously noted, companies that end up closing or getting acquired are often headed by a single individual.

4. It took Company B twice as long to secure their first round of funding, suggesting a slower initial impact in their industry.

5. Company A has raised over 17M USD, compared to Company B 10M USD.

6. Company A has completed one additional funding round compared to Company B.

7. Company A has attracted twice the number of investors as Company B, which could indicate a compelling business proposition.

Based on this data, the probability of success for these companies is:

Chance of Company A being successful
87%

Chance of Company B being successful
9%

As we can see, company A is closer to getting a 20+ million dollars funding round. While company B is close to being acquired or closed with just a 9% chance of being successful.

Final Thoughts:

In conclusion, the developed Logistic Regression model has proven to be a highly effective tool in assessing the performance and potential success of Latin American startups, as indicated by its impressive accuracy and AUC values. It has provided invaluable insights into the distinguishing factors between successful and unsuccessful companies, shedding light on the importance of factors like the timing of initial funding, the number of funding rounds, and the structure of the founding team.

Moreover, the model’s robust performance across both training and testing datasets underlines its reliability in real-world applications. The ability to accurately predict company success based on specific variables is of significant value to both investors and entrepreneurs. This model can help companies identify areas where they are underperforming and guide them to implement strategic changes to drive growth and success.

Limitations:

While the model has demonstrated robust performance, it does have some limitations.

One of the primary constraints is that the model relies heavily on the structure of the founding team, particularly whether the company has a solo founder or a team of co-founders. This may lead to some bias in the model’s predictions, as it places significant weight on this variable, potentially overlooking other critical factors.

Furthermore, the model is designed to analyze a specific set of variables, which may not encompass all factors that could influence a company’s success. Business outcomes can be influenced by a myriad of external factors, such as changes in market trends, economic fluctuations, and regulatory changes, which the model may not account for.

It is important to note that the model is based on historical data, which means its predictions are inherently rooted in past trends. While these trends can often provide a reliable guide, they do not guarantee future outcomes, as the business landscape is continually evolving.

Therefore, while the model provides valuable insights, it should not be the sole tool used in decision-making. It’s essential to use these predictions alongside other business analysis tools, expert opinion, and consideration of broader market and industry trends to make well-rounded decisions.